I think the inherit issue is the current “AI” is inherently non-deterministic, so it’s impossible to fix these issues totally. You can feed am AI all the data on how to not sound AI, but you need massive amounts of non-AI writing to reinforce that. With AI being so prevalent nowadays you can’t guarantee a dataset nowadays is AI free, so you get the old “garbage in garbage out” problem that AI companies cannot solve. I still think generative AI has it’s place as a tool, I use it for quick and dirty text manipulation, but it’s being applied to every problem we have like it’s a magic silver bullet. I’m ranting at this point and I’m going to stop here.
I honestly disagree that it has any use. Being a statistical model with high variance makes it a liability, no matter which task you use it for will produce worse results than a human being and will create new problems that didn’t exist before.
The high variance is why I only use it for dead simple tasks, e.g. “create and array of US states abbreviations in JavaScript”, otherwise I’m in full agreement with you. If you can’t verify the output is correct the it’s useless.
Wouldn’t that be slower to do, simply because checking it got all states, didn’t repeat any and didn’t make up any would be slower than copying a list from the web and quickly turning that into an array by hand with multiline cursors?
I think the best use is “making filler” so like in a game, having some deep background shit that no one looks at, or making a fake advertisement in a cyberpunk type game. Something to fill the world out that reduces the work of real artists if they choose to
FWIW, LLMs are deterministic. Usually the commercial front-ends don’t let you set the seed but behind the scenes the only reason the output changes each time it’s that the seed changes. If you set a fixed seed, input X always leads to output Y.
From the user perspective: no? I think they called that “temperature” and even setting that to 0 didn’t make the result the same the next day after cache cleared.
TBF OpenAI are a bunch of idiots running the world’s largest ponzi scheme. If DeepMind tried it and failed then…
Well I still wouldn’t be surprised, but at least it would be worth citing.
I think the inherit issue is the current “AI” is inherently non-deterministic, so it’s impossible to fix these issues totally. You can feed am AI all the data on how to not sound AI, but you need massive amounts of non-AI writing to reinforce that. With AI being so prevalent nowadays you can’t guarantee a dataset nowadays is AI free, so you get the old “garbage in garbage out” problem that AI companies cannot solve. I still think generative AI has it’s place as a tool, I use it for quick and dirty text manipulation, but it’s being applied to every problem we have like it’s a magic silver bullet. I’m ranting at this point and I’m going to stop here.
I honestly disagree that it has any use. Being a statistical model with high variance makes it a liability, no matter which task you use it for will produce worse results than a human being and will create new problems that didn’t exist before.
The high variance is why I only use it for dead simple tasks, e.g. “create and array of US states abbreviations in JavaScript”, otherwise I’m in full agreement with you. If you can’t verify the output is correct the it’s useless.
Wouldn’t that be slower to do, simply because checking it got all states, didn’t repeat any and didn’t make up any would be slower than copying a list from the web and quickly turning that into an array by hand with multiline cursors?
I think the best use is “making filler” so like in a game, having some deep background shit that no one looks at, or making a fake advertisement in a cyberpunk type game. Something to fill the world out that reduces the work of real artists if they choose to
If you can’t be bothered to write filler then it’s an insult for you to expect others to read it. You’re just wasting people’s time.
I guess the point is for people to not read the filler.
I think of the text that’s too small to read on a computer in the background. It’s nice that it’s slightly more real looking than a copy/paste screen.
Not even close to worth destroying the environment over, but it’s a neat use case to me
Lorem ipsum has been used in typesetting since the 60s. If it’s not meant to be read, it doesn’t matter if it’s lorem ipsum text.
Not trying to dogpile you, I just think even things that seem ‘useful’ for LLMs almost always have preexisting solutions that are decades old.
FWIW, LLMs are deterministic. Usually the commercial front-ends don’t let you set the seed but behind the scenes the only reason the output changes each time it’s that the seed changes. If you set a fixed seed, input X always leads to output Y.
From the user perspective: no? I think they called that “temperature” and even setting that to 0 didn’t make the result the same the next day after cache cleared.